def generate_uniform_dataset_reacher(env_class=None,
env_kwargs=None,
num_imgs=1000,
use_cached_dataset=False,
init_camera=None,
imsize=48,
show=False,
save_file_prefix=None,
env_id=None,
tag='',
dataset_path=None):
if dataset_path is not None:
dataset = load_local_or_remote_file(dataset_path)
return dataset
import gym
from gym.envs import registration
# trigger registration
import multiworld.envs.pygame
import multiworld.envs.mujoco
if not env_class or not env_kwargs:
env = gym.make(env_id)
else:
env = env_class(**env_kwargs)
env = ImageEnv(
env,
imsize,
init_camera=init_camera,
transpose=True,
normalize=True,
)
env.non_presampled_goal_img_is_garbage = True
if save_file_prefix is None and env_id is not None:
save_file_prefix = env_id
filename = "/tmp/{}_N{}_imsize{}uniform_images_{}.npy".format(
save_file_prefix,
str(num_imgs),
env.imsize,
tag,
)
if use_cached_dataset and osp.isfile(filename):
images = np.load(filename)
print("Loaded data from {}".format(filename))
return images
print('Sampling Uniform Dataset')
dataset = np.zeros((num_imgs, 3 * env.imsize**2), dtype=np.uint8)
for j in range(num_imgs):
obs = env.reset()
env.set_to_goal(env.get_goal())
img_f = env._get_flat_img()
if show:
img = img_f.reshape(3, env.imsize, env.imsize).transpose()
img = img[::-1, :, ::-1]
cv2.imshow('img', img)
cv2.waitKey(1)
print(j)
dataset[j, :] = unormalize_image(img_f)
np.save(filename, dataset)
print("Saving file to {}".format(filename))
return dataset
def generate_uniform_dataset_pick_and_place(env_class=None,
env_kwargs=None,
num_imgs=1000,
use_cached_dataset=False,
init_camera=None,
imsize=48,
save_file_prefix=None,
env_id=None,
tag='',
dataset_path=None):
if dataset_path is not None:
dataset = load_local_or_remote_file(dataset_path)
return dataset
import gym
from gym.envs import registration
# trigger registration
import multiworld.envs.pygame
import multiworld.envs.mujoco
if not env_class or not env_kwargs:
env = gym.make(env_id)
else:
env = env_class(**env_kwargs)
env = ImageEnv(
env,
imsize,
init_camera=init_camera,
transpose=True,
normalize=True,
)
env.non_presampled_goal_img_is_garbage = True
if save_file_prefix is None and env_id is not None:
save_file_prefix = env_id
filename = "/tmp/{}_N{}_imsize{}uniform_images_{}.npy".format(
save_file_prefix,
str(num_imgs),
env.imsize,
tag,
)
if use_cached_dataset and osp.isfile(filename):
images = np.load(filename)
print("Loaded data from {}".format(filename))
return images
print('Sampling Uniform Dataset')
dataset = unormalize_image(
get_image_presampled_goals(env, num_imgs)['image_desired_goal'])
np.save(filename, dataset)
print("Saving file to {}".format(filename))
return dataset
def HER_baseline_td3_experiment(variant):
import rlkit.torch.pytorch_util as ptu
from rlkit.data_management.obs_dict_replay_buffer import \
ObsDictRelabelingBuffer
from rlkit.exploration_strategies.base import (
PolicyWrappedWithExplorationStrategy)
from rlkit.torch.her.her_td3 import HerTd3
from rlkit.torch.networks import MergedCNN, CNNPolicy
import torch
from multiworld.core.image_env import ImageEnv
from rlkit.misc.asset_loader import load_local_or_remote_file
init_camera = variant.get("init_camera", None)
presample_goals = variant.get('presample_goals', False)
presampled_goals_path = get_presampled_goals_path(
variant.get('presampled_goals_path', None))
if 'env_id' in variant:
import gym
import multiworld
multiworld.register_all_envs()
env = gym.make(variant['env_id'])
else:
env = variant["env_class"](**variant['env_kwargs'])
image_env = ImageEnv(
env,
variant.get('imsize'),
reward_type='image_sparse',
init_camera=init_camera,
transpose=True,
normalize=True,
)
if presample_goals:
if presampled_goals_path is None:
image_env.non_presampled_goal_img_is_garbage = True
presampled_goals = variant['generate_goal_dataset_fctn'](
env=image_env, **variant['goal_generation_kwargs'])
else:
presampled_goals = load_local_or_remote_file(
presampled_goals_path).item()
del image_env
env = ImageEnv(
env,
variant.get('imsize'),
reward_type='image_distance',
init_camera=init_camera,
transpose=True,
normalize=True,
presampled_goals=presampled_goals,
)
else:
env = image_env
es = get_exploration_strategy(variant, env)
observation_key = variant.get('observation_key', 'image_observation')
desired_goal_key = variant.get('desired_goal_key', 'image_desired_goal')
achieved_goal_key = desired_goal_key.replace("desired", "achieved")
imsize = variant['imsize']
action_dim = env.action_space.low.size
qf1 = MergedCNN(input_width=imsize,
input_height=imsize,
output_size=1,
input_channels=3 * 2,
added_fc_input_size=action_dim,
**variant['cnn_params'])
qf2 = MergedCNN(input_width=imsize,
input_height=imsize,
output_size=1,
input_channels=3 * 2,
added_fc_input_size=action_dim,
**variant['cnn_params'])
policy = CNNPolicy(
input_width=imsize,
input_height=imsize,
added_fc_input_size=0,
output_size=action_dim,
input_channels=3 * 2,
output_activation=torch.tanh,
**variant['cnn_params'],
)
target_qf1 = MergedCNN(input_width=imsize,
input_height=imsize,
output_size=1,
input_channels=3 * 2,
added_fc_input_size=action_dim,
**variant['cnn_params'])
target_qf2 = MergedCNN(input_width=imsize,
input_height=imsize,
output_size=1,
input_channels=3 * 2,
added_fc_input_size=action_dim,
**variant['cnn_params'])
target_policy = CNNPolicy(
input_width=imsize,
input_height=imsize,
added_fc_input_size=0,
output_size=action_dim,
input_channels=3 * 2,
output_activation=torch.tanh,
**variant['cnn_params'],
)
exploration_policy = PolicyWrappedWithExplorationStrategy(
exploration_strategy=es,
policy=policy,
)
replay_buffer = ObsDictRelabelingBuffer(
env=env,
observation_key=observation_key,
desired_goal_key=desired_goal_key,
achieved_goal_key=achieved_goal_key,
**variant['replay_buffer_kwargs'])
algo_kwargs = variant['algo_kwargs']
algo_kwargs['replay_buffer'] = replay_buffer
base_kwargs = algo_kwargs['base_kwargs']
base_kwargs['training_env'] = env
base_kwargs['render'] = variant["render"]
base_kwargs['render_during_eval'] = variant["render"]
her_kwargs = algo_kwargs['her_kwargs']
her_kwargs['observation_key'] = observation_key
her_kwargs['desired_goal_key'] = desired_goal_key
algorithm = HerTd3(env,
qf1=qf1,
qf2=qf2,
policy=policy,
target_qf1=target_qf1,
target_qf2=target_qf2,
target_policy=target_policy,
exploration_policy=exploration_policy,
**variant['algo_kwargs'])
algorithm.to(ptu.device)
algorithm.train()
def get_envs(variant):
from multiworld.core.image_env import ImageEnv
from rlkit.envs.vae_wrapper import VAEWrappedEnv
from rlkit.util.io import load_local_or_remote_file
render = variant.get('render', False)
vae_path = variant.get("vae_path", None)
reward_params = variant.get("reward_params", dict())
init_camera = variant.get("init_camera", None)
do_state_exp = variant.get("do_state_exp", False)
presample_goals = variant.get('presample_goals', False)
presample_image_goals_only = variant.get('presample_image_goals_only',
False)
presampled_goals_path = variant.get('presampled_goals_path', None)
vae = load_local_or_remote_file(
vae_path) if type(vae_path) is str else vae_path
if 'env_id' in variant:
import gym
import multiworld
multiworld.register_all_envs()
env = gym.make(variant['env_id'])
else:
env = variant["env_class"](**variant['env_kwargs'])
if not do_state_exp:
if isinstance(env, ImageEnv):
image_env = env
else:
image_env = ImageEnv(
env,
variant.get('imsize'),
init_camera=init_camera,
transpose=True,
normalize=True,
)
if presample_goals:
"""
This will fail for online-parallel as presampled_goals will not be
serialized. Also don't use this for online-vae.
"""
if presampled_goals_path is None:
image_env.non_presampled_goal_img_is_garbage = True
vae_env = VAEWrappedEnv(image_env,
vae,
imsize=image_env.imsize,
decode_goals=render,
render_goals=render,
render_rollouts=render,
reward_params=reward_params,
**variant.get('vae_wrapped_env_kwargs',
{}))
presampled_goals = variant['generate_goal_dataset_fctn'](
env=vae_env,
env_id=variant.get('env_id', None),
**variant['goal_generation_kwargs'])
del vae_env
else:
presampled_goals = load_local_or_remote_file(
presampled_goals_path).item()
del image_env
image_env = ImageEnv(env,
variant.get('imsize'),
init_camera=init_camera,
transpose=True,
normalize=True,
presampled_goals=presampled_goals,
**variant.get('image_env_kwargs', {}))
vae_env = VAEWrappedEnv(image_env,
vae,
imsize=image_env.imsize,
decode_goals=render,
render_goals=render,
render_rollouts=render,
reward_params=reward_params,
presampled_goals=presampled_goals,
**variant.get('vae_wrapped_env_kwargs',
{}))
print("Presampling all goals only")
else:
vae_env = VAEWrappedEnv(image_env,
vae,
imsize=image_env.imsize,
decode_goals=render,
render_goals=render,
render_rollouts=render,
reward_params=reward_params,
**variant.get('vae_wrapped_env_kwargs',
{}))
if presample_image_goals_only:
presampled_goals = variant['generate_goal_dataset_fctn'](
image_env=vae_env.wrapped_env,
**variant['goal_generation_kwargs'])
image_env.set_presampled_goals(presampled_goals)
print("Presampling image goals only")
else:
print("Not using presampled goals")
env = vae_env
return env
def generate_vae_dataset(variant):
env_class = variant.get('env_class', None)
env_kwargs = variant.get('env_kwargs', None)
env_id = variant.get('env_id', None)
N = variant.get('N', 10000)
test_p = variant.get('test_p', 0.9)
use_cached = variant.get('use_cached', True)
imsize = variant.get('imsize', 84)
num_channels = variant.get('num_channels', 3)
show = variant.get('show', False)
init_camera = variant.get('init_camera', None)
dataset_path = variant.get('dataset_path', None)
oracle_dataset_using_set_to_goal = variant.get(
'oracle_dataset_using_set_to_goal', False)
random_rollout_data = variant.get('random_rollout_data', False)
random_and_oracle_policy_data = variant.get(
'random_and_oracle_policy_data', False)
random_and_oracle_policy_data_split = variant.get(
'random_and_oracle_policy_data_split', 0)
policy_file = variant.get('policy_file', None)
n_random_steps = variant.get('n_random_steps', 100)
vae_dataset_specific_env_kwargs = variant.get(
'vae_dataset_specific_env_kwargs', None)
save_file_prefix = variant.get('save_file_prefix', None)
non_presampled_goal_img_is_garbage = variant.get(
'non_presampled_goal_img_is_garbage', None)
tag = variant.get('tag', '')
from multiworld.core.image_env import ImageEnv, unormalize_image
import rlkit.torch.pytorch_util as ptu
info = {}
if dataset_path is not None:
dataset = load_local_or_remote_file(dataset_path)
N = dataset.shape[0]
else:
if env_kwargs is None:
env_kwargs = {}
if save_file_prefix is None:
save_file_prefix = env_id
if save_file_prefix is None:
save_file_prefix = env_class.__name__
filename = "/tmp/{}_N{}_{}_imsize{}_random_oracle_split_{}{}.npy".format(
save_file_prefix,
str(N),
init_camera.__name__ if init_camera else '',
imsize,
random_and_oracle_policy_data_split,
tag,
)
if use_cached and osp.isfile(filename):
dataset = np.load(filename)
print("loaded data from saved file", filename)
else:
now = time.time()
if env_id is not None:
import gym
import multiworld
multiworld.register_all_envs()
env = gym.make(env_id)
else:
if vae_dataset_specific_env_kwargs is None:
vae_dataset_specific_env_kwargs = {}
for key, val in env_kwargs.items():
if key not in vae_dataset_specific_env_kwargs:
vae_dataset_specific_env_kwargs[key] = val
env = env_class(**vae_dataset_specific_env_kwargs)
if not isinstance(env, ImageEnv):
env = ImageEnv(
env,
imsize,
init_camera=init_camera,
transpose=True,
normalize=True,
non_presampled_goal_img_is_garbage=
non_presampled_goal_img_is_garbage,
)
else:
imsize = env.imsize
env.non_presampled_goal_img_is_garbage = non_presampled_goal_img_is_garbage
env.reset()
info['env'] = env
if random_and_oracle_policy_data:
policy_file = load_local_or_remote_file(policy_file)
policy = policy_file['policy']
policy.to(ptu.device)
if random_rollout_data:
from rlkit.exploration_strategies.ou_strategy import OUStrategy
policy = OUStrategy(env.action_space)
dataset = np.zeros((N, imsize * imsize * num_channels),
dtype=np.uint8)
for i in range(N):
if random_and_oracle_policy_data:
num_random_steps = int(N *
random_and_oracle_policy_data_split)
if i < num_random_steps:
env.reset()
for _ in range(n_random_steps):
obs = env.step(env.action_space.sample())[0]
else:
obs = env.reset()
policy.reset()
for _ in range(n_random_steps):
policy_obs = np.hstack((
obs['state_observation'],
obs['state_desired_goal'],
))
action, _ = policy.get_action(policy_obs)
obs, _, _, _ = env.step(action)
elif oracle_dataset_using_set_to_goal:
print(i)
goal = env.sample_goal()
env.set_to_goal(goal)
obs = env._get_obs()
elif random_rollout_data:
if i % n_random_steps == 0:
g = dict(
state_desired_goal=env.sample_goal_for_rollout())
env.set_to_goal(g)
policy.reset()
# env.reset()
u = policy.get_action_from_raw_action(
env.action_space.sample())
obs = env.step(u)[0]
else:
env.reset()
for _ in range(n_random_steps):
obs = env.step(env.action_space.sample())[0]
img = obs['image_observation']
dataset[i, :] = unormalize_image(img)
if show:
img = img.reshape(3, imsize, imsize).transpose()
img = img[::-1, :, ::-1]
cv2.imshow('img', img)
cv2.waitKey(1)
# radius = input('waiting...')
print("done making training data", filename, time.time() - now)
np.save(filename, dataset)
n = int(N * test_p)
train_dataset = dataset[:n, :]
test_dataset = dataset[n:, :]
return train_dataset, test_dataset, info
def get_envs(variant):
render = variant.get('render', False)
vae_path = variant.get("vae_path", None)
reward_params = variant.get("reward_params", dict())
init_camera = variant.get("init_camera", None)
presample_goals = variant.get('presample_goals', False)
presample_image_goals_only = variant.get('presample_image_goals_only',
False)
presampled_goals_path = variant.get('presampled_goals_path', None)
vae = load_local_or_remote_file(
vae_path) if type(vae_path) is str else vae_path
if 'env_id' in variant:
env = gym.make(variant['env_id'])
else:
env = variant["env_class"](**variant['env_kwargs'])
if isinstance(env, ImageEnv):
image_env = env
else:
image_env = ImageEnv(
env,
variant.get('imsize'),
init_camera=init_camera,
transpose=True,
normalize=True,
)
if presample_goals:
"""
This will fail for online-parallel as presampled_goals will not be
serialized. Also don't use this for online-vae.
"""
if presampled_goals_path is None:
image_env.non_presampled_goal_img_is_garbage = True
vae_env = VAEWrappedEnv(image_env,
vae,
imsize=image_env.imsize,
decode_goals=render,
render_goals=render,
render_rollouts=render,
reward_params=reward_params,
**variant.get('vae_wrapped_env_kwargs',
{}))
presampled_goals = variant['generate_goal_dataset_fctn'](
env=vae_env,
env_id=variant.get('env_id', None),
**variant['goal_generation_kwargs'])
del vae_env
else:
presampled_goals = load_local_or_remote_file(
presampled_goals_path).item()
del image_env
image_env = ImageEnv(env,
variant.get('imsize'),
init_camera=init_camera,
transpose=True,
normalize=True,
presampled_goals=presampled_goals,
**variant.get('image_env_kwargs', {}))
vae_env = VAEWrappedEnv(image_env,
vae,
imsize=image_env.imsize,
decode_goals=render,
render_goals=render,
render_rollouts=render,
reward_params=reward_params,
presampled_goals=presampled_goals,
**variant.get('vae_wrapped_env_kwargs', {}))
print("Presampling all goals only")
else:
vae_env = VAEWrappedEnv(image_env,
vae,
imsize=image_env.imsize,
decode_goals=render,
render_goals=render,
render_rollouts=render,
reward_params=reward_params,
**variant.get('vae_wrapped_env_kwargs', {}))
if presample_image_goals_only:
presampled_goals = variant['generate_goal_dataset_fctn'](
image_env=vae_env.wrapped_env,
**variant['goal_generation_kwargs'])
image_env.set_presampled_goals(presampled_goals)
print("Presampling image goals only")
else:
print("Not using presampled goals")
env = vae_env
training_mode = variant.get("training_mode", "train")
testing_mode = variant.get("testing_mode", "test")
env.add_mode('eval', testing_mode)
env.add_mode('train', training_mode)
env.add_mode('relabeling', training_mode)
# relabeling_env.disable_render()
env.add_mode("video_vae", 'video_vae')
env.add_mode("video_env", 'video_env')
return env
def get_envs(variant):
from multiworld.core.image_env import ImageEnv
from railrl.envs.vae_wrappers import VAEWrappedEnv
from railrl.misc.asset_loader import load_local_or_remote_file
render = variant.get('render', False)
vae_path = variant.get("vae_path", None)
reproj_vae_path = variant.get("reproj_vae_path", None)
ckpt = variant.get("ckpt", None)
reward_params = variant.get("reward_params", dict())
init_camera = variant.get("init_camera", None)
do_state_exp = variant.get("do_state_exp", False)
presample_goals = variant.get('presample_goals', False)
presample_image_goals_only = variant.get('presample_image_goals_only', False)
presampled_goals_path = variant.get('presampled_goals_path', None)
if not do_state_exp and type(ckpt) is str:
vae = load_local_or_remote_file(osp.join(ckpt, 'vae.pkl'))
if vae is not None:
from railrl.core import logger
logger.save_extra_data(vae, 'vae.pkl', mode='pickle')
else:
vae = None
if vae is None and type(vae_path) is str:
vae = load_local_or_remote_file(osp.join(vae_path, 'vae_params.pkl'))
from railrl.core import logger
logger.save_extra_data(vae, 'vae.pkl', mode='pickle')
elif vae is None:
vae = vae_path
if type(vae) is str:
vae = load_local_or_remote_file(vae)
else:
vae = vae
if type(reproj_vae_path) is str:
reproj_vae = load_local_or_remote_file(osp.join(reproj_vae_path, 'vae_params.pkl'))
else:
reproj_vae = None
if 'env_id' in variant:
import gym
# trigger registration
env = gym.make(variant['env_id'])
else:
env = variant["env_class"](**variant['env_kwargs'])
if not do_state_exp:
if isinstance(env, ImageEnv):
image_env = env
else:
image_env = ImageEnv(
env,
variant.get('imsize'),
init_camera=init_camera,
transpose=True,
normalize=True,
)
vae_env = VAEWrappedEnv(
image_env,
vae,
imsize=image_env.imsize,
decode_goals=render,
render_goals=render,
render_rollouts=render,
reward_params=reward_params,
reproj_vae=reproj_vae,
**variant.get('vae_wrapped_env_kwargs', {})
)
if presample_goals:
"""
This will fail for online-parallel as presampled_goals will not be
serialized. Also don't use this for online-vae.
"""
if presampled_goals_path is None:
image_env.non_presampled_goal_img_is_garbage = True
presampled_goals = variant['generate_goal_dataset_fctn'](
image_env=image_env,
**variant['goal_generation_kwargs']
)
else:
presampled_goals = load_local_or_remote_file(
presampled_goals_path
).item()
presampled_goals = {
'state_desired_goal': presampled_goals['next_obs_state'],
'image_desired_goal': presampled_goals['next_obs'],
}
image_env.set_presampled_goals(presampled_goals)
vae_env.set_presampled_goals(presampled_goals)
print("Presampling all goals")
else:
if presample_image_goals_only:
presampled_goals = variant['generate_goal_dataset_fctn'](
image_env=vae_env.wrapped_env,
**variant['goal_generation_kwargs']
)
image_env.set_presampled_goals(presampled_goals)
print("Presampling image goals only")
else:
print("Not using presampled goals")
env = vae_env
if not do_state_exp:
training_mode = variant.get("training_mode", "train")
testing_mode = variant.get("testing_mode", "test")
env.add_mode('eval', testing_mode)
env.add_mode('train', training_mode)
env.add_mode('relabeling', training_mode)
# relabeling_env.disable_render()
env.add_mode("video_vae", 'video_vae')
env.add_mode("video_env", 'video_env')
return env
def generate_vae_dataset(variant):
print(variant)
from tqdm import tqdm
env_class = variant.get('env_class', None)
env_kwargs = variant.get('env_kwargs', None)
env_id = variant.get('env_id', None)
N = variant.get('N', 10000)
batch_size = variant.get('batch_size', 128)
test_p = variant.get('test_p', 0.9)
use_cached = variant.get('use_cached', True)
imsize = variant.get('imsize', 84)
num_channels = variant.get('num_channels', 3)
show = variant.get('show', False)
init_camera = variant.get('init_camera', None)
dataset_path = variant.get('dataset_path', None)
augment_data = variant.get('augment_data', False)
data_filter_fn = variant.get('data_filter_fn', lambda x: x)
delete_after_loading = variant.get('delete_after_loading', False)
oracle_dataset_using_set_to_goal = variant.get(
'oracle_dataset_using_set_to_goal', False)
random_rollout_data = variant.get('random_rollout_data', False)
random_rollout_data_set_to_goal = variant.get(
'random_rollout_data_set_to_goal', True)
random_and_oracle_policy_data = variant.get(
'random_and_oracle_policy_data', False)
random_and_oracle_policy_data_split = variant.get(
'random_and_oracle_policy_data_split', 0)
policy_file = variant.get('policy_file', None)
n_random_steps = variant.get('n_random_steps', 100)
vae_dataset_specific_env_kwargs = variant.get(
'vae_dataset_specific_env_kwargs', None)
save_file_prefix = variant.get('save_file_prefix', None)
non_presampled_goal_img_is_garbage = variant.get(
'non_presampled_goal_img_is_garbage', None)
conditional_vae_dataset = variant.get('conditional_vae_dataset', False)
use_env_labels = variant.get('use_env_labels', False)
use_linear_dynamics = variant.get('use_linear_dynamics', False)
enviorment_dataset = variant.get('enviorment_dataset', False)
save_trajectories = variant.get('save_trajectories', False)
save_trajectories = save_trajectories or use_linear_dynamics or conditional_vae_dataset
tag = variant.get('tag', '')
assert N % n_random_steps == 0, "Fix N/horizon or dataset generation will fail"
from multiworld.core.image_env import ImageEnv, unormalize_image
import rlkit.torch.pytorch_util as ptu
from rlkit.util.io import load_local_or_remote_file
from rlkit.data_management.dataset import (
TrajectoryDataset, ImageObservationDataset, InitialObservationDataset,
EnvironmentDataset, ConditionalDynamicsDataset,
InitialObservationNumpyDataset, InfiniteBatchLoader,
InitialObservationNumpyJitteringDataset)
info = {}
use_test_dataset = False
if dataset_path is not None:
if type(dataset_path) == str:
dataset = load_local_or_remote_file(
dataset_path, delete_after_loading=delete_after_loading)
dataset = dataset.item()
N = dataset['observations'].shape[0] * dataset[
'observations'].shape[1]
n_random_steps = dataset['observations'].shape[1]
if isinstance(dataset_path, list):
dataset = concatenate_datasets(dataset_path)
N = dataset['observations'].shape[0] * dataset[
'observations'].shape[1]
n_random_steps = dataset['observations'].shape[1]
if isinstance(dataset_path, dict):
if type(dataset_path['train']) == str:
dataset = load_local_or_remote_file(
dataset_path['train'],
delete_after_loading=delete_after_loading)
dataset = dataset.item()
elif isinstance(dataset_path['train'], list):
dataset = concatenate_datasets(dataset_path['train'])
if type(dataset_path['test']) == str:
test_dataset = load_local_or_remote_file(
dataset_path['test'],
delete_after_loading=delete_after_loading)
test_dataset = test_dataset.item()
elif isinstance(dataset_path['test'], list):
test_dataset = concatenate_datasets(dataset_path['test'])
N = dataset['observations'].shape[0] * dataset[
'observations'].shape[1]
n_random_steps = dataset['observations'].shape[1]
use_test_dataset = True
else:
if env_kwargs is None:
env_kwargs = {}
if save_file_prefix is None:
save_file_prefix = env_id
if save_file_prefix is None:
save_file_prefix = env_class.__name__
filename = "/tmp/{}_N{}_{}_imsize{}_random_oracle_split_{}{}.npy".format(
save_file_prefix,
str(N),
init_camera.__name__
if init_camera and hasattr(init_camera, '__name__') else '',
imsize,
random_and_oracle_policy_data_split,
tag,
)
if use_cached and osp.isfile(filename):
dataset = load_local_or_remote_file(
filename, delete_after_loading=delete_after_loading)
if conditional_vae_dataset:
dataset = dataset.item()
print("loaded data from saved file", filename)
else:
now = time.time()
if env_id is not None:
import gym
import multiworld
multiworld.register_all_envs()
env = gym.make(env_id)
else:
if vae_dataset_specific_env_kwargs is None:
vae_dataset_specific_env_kwargs = {}
for key, val in env_kwargs.items():
if key not in vae_dataset_specific_env_kwargs:
vae_dataset_specific_env_kwargs[key] = val
env = env_class(**vae_dataset_specific_env_kwargs)
if not isinstance(env, ImageEnv):
env = ImageEnv(
env,
imsize,
init_camera=init_camera,
transpose=True,
normalize=True,
non_presampled_goal_img_is_garbage=
non_presampled_goal_img_is_garbage,
)
else:
imsize = env.imsize
env.non_presampled_goal_img_is_garbage = non_presampled_goal_img_is_garbage
env.reset()
info['env'] = env
if random_and_oracle_policy_data:
policy_file = load_local_or_remote_file(policy_file)
policy = policy_file['policy']
policy.to(ptu.device)
if random_rollout_data:
from rlkit.exploration_strategies.ou_strategy import OUStrategy
policy = OUStrategy(env.action_space)
if save_trajectories:
dataset = {
'observations':
np.zeros((N // n_random_steps, n_random_steps,
imsize * imsize * num_channels),
dtype=np.uint8),
'actions':
np.zeros((N // n_random_steps, n_random_steps,
env.action_space.shape[0]),
dtype=np.float),
'env':
np.zeros(
(N // n_random_steps, imsize * imsize * num_channels),
dtype=np.uint8),
}
else:
dataset = np.zeros((N, imsize * imsize * num_channels),
dtype=np.uint8)
labels = []
for i in tqdm(range(N)):
if random_and_oracle_policy_data:
num_random_steps = int(N *
random_and_oracle_policy_data_split)
if i < num_random_steps:
env.reset()
for _ in range(n_random_steps):
obs = env.step(env.action_space.sample())[0]
else:
obs = env.reset()
policy.reset()
for _ in range(n_random_steps):
policy_obs = np.hstack((
obs['state_observation'],
obs['state_desired_goal'],
))
action, _ = policy.get_action(policy_obs)
obs, _, _, _ = env.step(action)
elif random_rollout_data: #ADD DATA WHERE JUST PUCK MOVES
if i % n_random_steps == 0:
env.reset()
policy.reset()
env_img = env._get_obs()['image_observation']
if random_rollout_data_set_to_goal:
env.set_to_goal(env.get_goal())
obs = env._get_obs()
u = policy.get_action_from_raw_action(
env.action_space.sample())
env.step(u)
elif oracle_dataset_using_set_to_goal:
print(i)
goal = env.sample_goal()
env.set_to_goal(goal)
obs = env._get_obs()
else:
env.reset()
for _ in range(n_random_steps):
obs = env.step(env.action_space.sample())[0]
img = obs['image_observation']
if use_env_labels:
labels.append(obs['label'])
if save_trajectories:
dataset['observations'][
i // n_random_steps,
i % n_random_steps, :] = unormalize_image(img)
dataset['actions'][i // n_random_steps,
i % n_random_steps, :] = u
dataset['env'][i // n_random_steps, :] = unormalize_image(
env_img)
else:
dataset[i, :] = unormalize_image(img)
if show:
img = img.reshape(3, imsize, imsize).transpose()
img = img[::-1, :, ::-1]
cv2.imshow('img', img)
cv2.waitKey(1)
# radius = input('waiting...')
print("done making training data", filename, time.time() - now)
np.save(filename, dataset)
#np.save(filename[:-4] + 'labels.npy', np.array(labels))
info['train_labels'] = []
info['test_labels'] = []
dataset = data_filter_fn(dataset)
if use_linear_dynamics and conditional_vae_dataset:
num_trajectories = N // n_random_steps
n = int(num_trajectories * test_p)
train_dataset = ConditionalDynamicsDataset({
'observations':
dataset['observations'][:n, :, :],
'actions':
dataset['actions'][:n, :, :],
'env':
dataset['env'][:n, :]
})
test_dataset = ConditionalDynamicsDataset({
'observations':
dataset['observations'][n:, :, :],
'actions':
dataset['actions'][n:, :, :],
'env':
dataset['env'][n:, :]
})
num_trajectories = N // n_random_steps
n = int(num_trajectories * test_p)
indices = np.arange(num_trajectories)
np.random.shuffle(indices)
train_i, test_i = indices[:n], indices[n:]
try:
train_dataset = ConditionalDynamicsDataset({
'observations':
dataset['observations'][train_i, :, :],
'actions':
dataset['actions'][train_i, :, :],
'env':
dataset['env'][train_i, :]
})
test_dataset = ConditionalDynamicsDataset({
'observations':
dataset['observations'][test_i, :, :],
'actions':
dataset['actions'][test_i, :, :],
'env':
dataset['env'][test_i, :]
})
except:
train_dataset = ConditionalDynamicsDataset({
'observations':
dataset['observations'][train_i, :, :],
'actions':
dataset['actions'][train_i, :, :],
})
test_dataset = ConditionalDynamicsDataset({
'observations':
dataset['observations'][test_i, :, :],
'actions':
dataset['actions'][test_i, :, :],
})
elif use_linear_dynamics:
num_trajectories = N // n_random_steps
n = int(num_trajectories * test_p)
train_dataset = TrajectoryDataset({
'observations':
dataset['observations'][:n, :, :],
'actions':
dataset['actions'][:n, :, :]
})
test_dataset = TrajectoryDataset({
'observations':
dataset['observations'][n:, :, :],
'actions':
dataset['actions'][n:, :, :]
})
elif enviorment_dataset:
n = int(n_random_steps * test_p)
train_dataset = EnvironmentDataset({
'observations':
dataset['observations'][:, :n, :],
})
test_dataset = EnvironmentDataset({
'observations':
dataset['observations'][:, n:, :],
})
elif conditional_vae_dataset:
num_trajectories = N // n_random_steps
n = int(num_trajectories * test_p)
indices = np.arange(num_trajectories)
np.random.shuffle(indices)
train_i, test_i = indices[:n], indices[n:]
if augment_data:
dataset_class = InitialObservationNumpyJitteringDataset
else:
dataset_class = InitialObservationNumpyDataset
if 'env' not in dataset:
dataset['env'] = dataset['observations'][:, 0]
if use_test_dataset and ('env' not in test_dataset):
test_dataset['env'] = test_dataset['observations'][:, 0]
if use_test_dataset:
train_dataset = dataset_class({
'observations':
dataset['observations'],
'env':
dataset['env']
})
test_dataset = dataset_class({
'observations':
test_dataset['observations'],
'env':
test_dataset['env']
})
else:
train_dataset = dataset_class({
'observations':
dataset['observations'][train_i, :, :],
'env':
dataset['env'][train_i, :]
})
test_dataset = dataset_class({
'observations':
dataset['observations'][test_i, :, :],
'env':
dataset['env'][test_i, :]
})
train_batch_loader_kwargs = variant.get(
'train_batch_loader_kwargs',
dict(
batch_size=batch_size,
num_workers=0,
))
test_batch_loader_kwargs = variant.get(
'test_batch_loader_kwargs',
dict(
batch_size=batch_size,
num_workers=0,
))
train_data_loader = data.DataLoader(train_dataset,
shuffle=True,
drop_last=True,
**train_batch_loader_kwargs)
test_data_loader = data.DataLoader(test_dataset,
shuffle=True,
drop_last=True,
**test_batch_loader_kwargs)
train_dataset = InfiniteBatchLoader(train_data_loader)
test_dataset = InfiniteBatchLoader(test_data_loader)
else:
n = int(N * test_p)
train_dataset = ImageObservationDataset(dataset[:n, :])
test_dataset = ImageObservationDataset(dataset[n:, :])
return train_dataset, test_dataset, info
def getdata(variant):
skewfit_variant = variant['skewfit_variant']
print('-------------------------------')
skewfit_preprocess_variant(skewfit_variant)
skewfit_variant['render'] = True
vae_environment = get_envs(skewfit_variant)
print('done loading vae_env')
env_class = variant.get('env_class', None)
env_kwargs = variant.get('env_kwargs', None)
env_id = variant.get('env_id', None)
N = variant.get('N', 10000)
test_p = variant.get('test_p', 0.9)
use_cached = variant.get('use_cached', True)
imsize = variant.get('imsize', 84)
num_channels = variant.get('num_channels', 3)
show = variant.get('show', False)
init_camera = variant.get('init_camera', None)
dataset_path = variant.get('dataset_path', None)
oracle_dataset_using_set_to_goal = variant.get(
'oracle_dataset_using_set_to_goal', False)
random_rollout_data = variant.get('random_rollout_data', False)
random_and_oracle_policy_data = variant.get(
'random_and_oracle_policy_data', False)
random_and_oracle_policy_data_split = variant.get(
'random_and_oracle_policy_data_split', 0)
policy_file = variant.get('policy_file', None)
n_random_steps = variant.get('n_random_steps', 100)
vae_dataset_specific_env_kwargs = variant.get(
'vae_dataset_specific_env_kwargs', None)
save_file_prefix = variant.get('save_file_prefix', None)
non_presampled_goal_img_is_garbage = variant.get(
'non_presampled_goal_img_is_garbage', None)
tag = variant.get('tag', '')
from multiworld.core.image_env import ImageEnv, unormalize_image
import rlkit.torch.pytorch_util as ptu
info = {}
if dataset_path is not None:
dataset = load_local_or_remote_file(dataset_path)
N = dataset.shape[0]
else:
if env_kwargs is None:
env_kwargs = {}
if save_file_prefix is None:
save_file_prefix = env_id
if save_file_prefix is None:
save_file_prefix = env_class.__name__
filename = "/tmp/{}_N{}_{}_imsize{}_random_oracle_split_{}{}.npy".format(
save_file_prefix,
str(N),
init_camera.__name__ if init_camera else '',
imsize,
random_and_oracle_policy_data_split,
tag,
)
if True:
now = time.time()
if env_id is not None:
import gym
import multiworld
multiworld.register_all_envs()
env = gym.make(env_id)
else:
if vae_dataset_specific_env_kwargs is None:
vae_dataset_specific_env_kwargs = {}
for key, val in env_kwargs.items():
if key not in vae_dataset_specific_env_kwargs:
vae_dataset_specific_env_kwargs[key] = val
env = env_class(**vae_dataset_specific_env_kwargs)
if not isinstance(env, ImageEnv):
print("using(ImageEnv)")
env = ImageEnv(
env,
imsize,
init_camera=init_camera,
transpose=True,
normalize=True,
non_presampled_goal_img_is_garbage=
non_presampled_goal_img_is_garbage,
)
else:
imsize = env.imsize
env.non_presampled_goal_img_is_garbage = non_presampled_goal_img_is_garbage
env.reset()
info['env'] = env
if random_and_oracle_policy_data:
policy_file = load_local_or_remote_file(policy_file)
policy = policy_file['policy']
policy.to(ptu.device)
if random_rollout_data:
from rlkit.exploration_strategies.ou_strategy import OUStrategy
policy = OUStrategy(env.action_space)
dataset = np.zeros((N, imsize * imsize * num_channels),
dtype=np.uint8)
for i in range(10):
NP = []
if True:
print(i)
#print('th step')
goal = env.sample_goal()
# print("goal___________________________")
# print(goal)
# print("goal___________________________")
env.set_to_goal(goal)
obs = env._get_obs()
#img = img.reshape(3, imsize, imsize).transpose()
# img = img[::-1, :, ::-1]
# cv2.imshow('img', img)
# cv2.waitKey(1)
img_1 = obs['image_observation']
img_1 = img_1.reshape(3, imsize, imsize).transpose()
NP.append(img_1)
if i % 3 == 0:
cv2.imshow('img1', img_1)
cv2.waitKey(1)
#img_1_reconstruct = vae_environment._reconstruct_img(obs['image_observation']).transpose()
encoded_1 = vae_environment._get_encoded(
obs['image_observation'])
print(encoded_1)
NP.append(encoded_1)
img_1_reconstruct = vae_environment._get_img(
encoded_1).transpose()
NP.append(img_1_reconstruct)
#dataset[i, :] = unormalize_image(img)
# img_1 = img_1.reshape(3, imsize, imsize).transpose()
if i % 3 == 0:
cv2.imshow('img1_reconstruction', img_1_reconstruct)
cv2.waitKey(1)
env.reset()
instr = env.generate_new_state(goal)
if i % 3 == 0:
print(instr)
obs = env._get_obs()
# obs = env._get_obs()
img_2 = obs['image_observation']
img_2 = img_2.reshape(3, imsize, imsize).transpose()
NP.append(img_2)
if i % 3 == 0:
cv2.imshow('img2', img_2)
cv2.waitKey(1)
#img_2_reconstruct = vae_environment._reconstruct_img(obs['image_observation']).transpose()
encoded_2 = vae_environment._get_encoded(
obs['image_observation'])
NP.append(encoded_2)
img_2_reconstruct = vae_environment._get_img(
encoded_2).transpose()
NP.append(img_2_reconstruct)
NP.append(instr)
# img_2 = img_2.reshape(3, imsize, imsize).transpose()
if i % 3 == 0:
cv2.imshow('img2_reconstruct', img_2_reconstruct)
cv2.waitKey(1)
NP = np.array(NP)
idx = str(i)
name = "/home/xiaomin/Downloads/IFIG_DATA_1/" + idx + ".npy"
np.save(open(name, 'wb'), NP)
# radius = input('waiting...')
# #get the in between functions
import dill
import pickle
get_encoded = dill.dumps(vae_environment._get_encoded)
with open(
"/home/xiaomin/Downloads/IFIG_encoder_decoder/get_encoded_1000_epochs_one_puck.txt",
"wb") as fp:
pickle.dump(get_encoded, fp)
with open(
"/home/xiaomin/Downloads/IFIG_encoder_decoder/get_encoded_1000_epochs_one_puck.txt",
"rb") as fp:
b = pickle.load(fp)
func_get_encoded = dill.loads(b)
encoded = func_get_encoded(obs['image_observation'])
print(encoded)
print('------------------------------')
get_img = dill.dumps(vae_environment._get_img)
with open(
"/home/xiaomin/Downloads/IFIG_encoder_decoder/get_img_1000_epochs_one_puck.txt",
"wb") as fp:
pickle.dump(get_img, fp)
with open(
"/home/xiaomin/Downloads/IFIG_encoder_decoder/get_img_1000_epochs_one_puck.txt",
"rb") as fp:
c = pickle.load(fp)
func_get_img = dill.loads(c)
img_1_reconstruct = func_get_img(encoded).transpose()
print(img_1_reconstruct)
#dataset[i, :] = unormalize_image(img)
# img_1 = img_1.reshape(3, imsize, imsize).transpose()
cv2.imshow('test', img_1_reconstruct)
cv2.waitKey(0)
print("done making training data", filename, time.time() - now)
np.save(filename, dataset)
def generate_vae_dataset(variant):
"""
If not provided a pre-train vae dataset generation function, this function will be used to collect
the dataset for training vae.
"""
import rlkit.torch.pytorch_util as ptu
import gym
import multiworld
multiworld.register_all_envs()
print("generating vae dataset with original images")
env_class = variant.get('env_class', None)
env_kwargs = variant.get('env_kwargs', None)
env_id = variant.get('env_id', None)
N = variant.get('N', 10000)
test_p = variant.get('test_p', 0.9)
use_cached = variant.get('use_cached', True)
imsize = variant.get('imsize', 84)
num_channels = variant.get('num_channels', 3)
show = variant.get('show', False)
init_camera = variant.get('init_camera', None)
dataset_path = variant.get('dataset_path', None)
oracle_dataset_using_set_to_goal = variant.get(
'oracle_dataset_using_set_to_goal', False)
random_rollout_data = variant.get('random_rollout_data', False)
random_and_oracle_policy_data = variant.get(
'random_and_oracle_policy_data', False)
random_and_oracle_policy_data_split = variant.get(
'random_and_oracle_policy_data_split', 0)
policy_file = variant.get('policy_file', None)
n_random_steps = variant.get('n_random_steps', 100)
vae_dataset_specific_env_kwargs = variant.get(
'vae_dataset_specific_env_kwargs', None)
save_file_prefix = variant.get('save_file_prefix', None)
non_presampled_goal_img_is_garbage = variant.get(
'non_presampled_goal_img_is_garbage', None)
tag = variant.get('tag', '')
info = {}
if dataset_path is not None:
print('load vae training dataset from: ', dataset_path)
pjhome = os.environ['PJHOME']
dataset = np.load(osp.join(pjhome, dataset_path),
allow_pickle=True).item()
if isinstance(dataset, dict):
dataset = dataset['image_desired_goal']
dataset = unormalize_image(dataset)
N = dataset.shape[0]
else:
if env_kwargs is None:
env_kwargs = {}
if save_file_prefix is None:
save_file_prefix = env_id
if save_file_prefix is None:
save_file_prefix = env_class.__name__
filename = "/tmp/{}_N{}_{}_imsize{}_random_oracle_split_{}{}.npy".format(
save_file_prefix,
str(N),
init_camera.__name__ if init_camera else '',
imsize,
random_and_oracle_policy_data_split,
tag,
)
if use_cached and osp.isfile(filename):
dataset = np.load(filename)
print("loaded data from saved file", filename)
else:
now = time.time()
if env_id is not None:
import gym
import multiworld
multiworld.register_all_envs()
env = gym.make(env_id)
else:
if vae_dataset_specific_env_kwargs is None:
vae_dataset_specific_env_kwargs = {}
for key, val in env_kwargs.items():
if key not in vae_dataset_specific_env_kwargs:
vae_dataset_specific_env_kwargs[key] = val
env = env_class(**vae_dataset_specific_env_kwargs)
if not isinstance(env, ImageEnv):
env = ImageEnv(
env,
imsize,
init_camera=init_camera,
transpose=True,
normalize=True,
non_presampled_goal_img_is_garbage=
non_presampled_goal_img_is_garbage,
)
else:
imsize = env.imsize
env.non_presampled_goal_img_is_garbage = non_presampled_goal_img_is_garbage
env.reset()
info['env'] = env
if random_and_oracle_policy_data:
policy_file = load_local_or_remote_file(policy_file)
policy = policy_file['policy']
policy.to(ptu.device)
if random_rollout_data:
from rlkit.exploration_strategies.ou_strategy import OUStrategy
policy = OUStrategy(env.action_space)
dataset = np.zeros((N, imsize * imsize * num_channels),
dtype=np.uint8)
for i in range(N):
if random_and_oracle_policy_data:
num_random_steps = int(N *
random_and_oracle_policy_data_split)
if i < num_random_steps:
env.reset()
for _ in range(n_random_steps):
obs = env.step(env.action_space.sample())[0]
else:
obs = env.reset()
policy.reset()
for _ in range(n_random_steps):
policy_obs = np.hstack((
obs['state_observation'],
obs['state_desired_goal'],
))
action, _ = policy.get_action(policy_obs)
obs, _, _, _ = env.step(action)
elif oracle_dataset_using_set_to_goal:
print(i)
goal = env.sample_goal()
env.set_to_goal(goal)
obs = env._get_obs()
elif random_rollout_data:
if i % n_random_steps == 0:
g = dict(
state_desired_goal=env.sample_goal_for_rollout())
env.set_to_goal(g)
policy.reset()
# env.reset()
u = policy.get_action_from_raw_action(
env.action_space.sample())
obs = env.step(u)[0]
else:
print("using totally random rollouts")
for _ in range(n_random_steps):
obs = env.step(env.action_space.sample())[0]
img = obs[
'image_observation'] # NOTE yufei: this is already normalized image, of detype np.float64.
dataset[i, :] = unormalize_image(img)
np.save(filename, dataset)
n = int(N * test_p)
train_dataset = dataset[:n, :]
test_dataset = dataset[n:, :]
return train_dataset, test_dataset, info
def generate_uniform_dataset_door(num_imgs=1000,
use_cached_dataset=False,
init_camera=None,
imsize=48,
policy_file=None,
show=False,
path_length=100,
save_file_prefix=None,
env_id=None,
tag='',
dataset_path=None):
if dataset_path is not None:
dataset = load_local_or_remote_file(dataset_path)
return dataset
import gym
from gym.envs import registration
# trigger registration
import multiworld.envs.pygame
import multiworld.envs.mujoco
env = gym.make(env_id)
env = ImageEnv(
env,
imsize,
init_camera=init_camera,
transpose=True,
normalize=True,
)
env.non_presampled_goal_img_is_garbage = True
if save_file_prefix is None and env_id is not None:
save_file_prefix = env_id
filename = "/tmp/{}_N{}_imsize{}uniform_images_{}.npy".format(
save_file_prefix,
str(num_imgs),
env.imsize,
tag,
)
if use_cached_dataset and osp.isfile(filename):
images = np.load(filename)
print("Loaded data from {}".format(filename))
return images
policy_file = load_local_or_remote_file(policy_file)
policy = policy_file['policy']
policy.to(ptu.device)
print('Sampling Uniform Dataset')
dataset = np.zeros((num_imgs, 3 * env.imsize**2), dtype=np.uint8)
for j in range(num_imgs):
obs = env.reset()
policy.reset()
for i in range(path_length):
policy_obs = np.hstack((
obs['state_observation'],
obs['state_desired_goal'],
))
action, _ = policy.get_action(policy_obs)
obs, _, _, _ = env.step(action)
img_f = obs['image_observation']
if show:
img = obs['image_observation']
img = img.reshape(3, env.imsize, env.imsize).transpose()
img = img[::-1, :, ::-1]
cv2.imshow('img', img)
cv2.waitKey(1)
print(j)
dataset[j, :] = unormalize_image(img_f)
temp = env.reset_free
env.reset_free = True
env.reset()
env.reset_free = temp
np.save(filename, dataset)
print("Saving file to {}".format(filename))
return dataset
def generate_vae_dataset(variant):
import cv2
env_class = variant.get('env_class', None)
env_kwargs = variant.get('env_kwargs', None)
env_id = variant.get('env_id', None)
N = variant.get('N', 10000)
use_images = variant.get('use_images', True)
imsize = variant.get('imsize', 84)
show = variant.get('show', False)
init_camera = variant.get('init_camera', None)
oracle_dataset = variant.get('oracle_dataset', False)
if 'n_random_steps' in variant:
n_random_steps = variant['n_random_steps']
else:
if oracle_dataset:
n_random_steps = 3
else:
n_random_steps = 100
vae_dataset_specific_env_kwargs = variant.get(
'vae_dataset_specific_env_kwargs', None)
non_presampled_goal_img_is_garbage = variant.get(
'non_presampled_goal_img_is_garbage', None)
from multiworld.core.image_env import ImageEnv, unormalize_image
info = {}
from railrl.core import logger
logdir = logger.get_snapshot_dir()
filename = osp.join(logdir, "vae_dataset.npy")
now = time.time()
if env_id is not None:
import gym
env = gym.make(env_id)
else:
if vae_dataset_specific_env_kwargs is None:
vae_dataset_specific_env_kwargs = {}
for key, val in env_kwargs.items():
if key not in vae_dataset_specific_env_kwargs:
vae_dataset_specific_env_kwargs[key] = val
env = env_class(**vae_dataset_specific_env_kwargs)
if not isinstance(env, ImageEnv):
env = ImageEnv(
env,
imsize,
init_camera=init_camera,
transpose=True,
normalize=True,
non_presampled_goal_img_is_garbage=
non_presampled_goal_img_is_garbage,
)
else:
imsize = env.imsize
env.non_presampled_goal_img_is_garbage = non_presampled_goal_img_is_garbage
env.reset()
info['env'] = env
if use_images:
data_size = len(env.observation_space.spaces['image_observation'].low)
dtype = np.uint8
else:
data_size = len(env.observation_space.spaces['state_observation'].low)
dtype = np.float32
state_size = len(env.observation_space.spaces['state_observation'].low)
dataset = {
'obs': np.zeros((N, data_size), dtype=dtype),
'actions': np.zeros((N, len(env.action_space.low)), dtype=np.float32),
'next_obs': np.zeros((N, data_size), dtype=dtype),
'obs_state': np.zeros((N, state_size), dtype=np.float32),
'next_obs_state': np.zeros((N, state_size), dtype=np.float32),
}
for i in range(N):
if i % (N / 50) == 0:
print(i)
if oracle_dataset:
if i % 100 == 0:
env.reset()
goal = env.sample_goal()
env.set_to_goal(goal)
for _ in range(n_random_steps):
env.step(env.action_space.sample())
else:
env.reset()
for _ in range(n_random_steps):
env.step(env.action_space.sample())
obs = env._get_obs()
if use_images:
dataset['obs'][i, :] = unormalize_image(obs['image_observation'])
else:
dataset['obs'][i, :] = obs['state_observation']
dataset['obs_state'][i, :] = obs['state_observation']
action = env.action_space.sample()
dataset['actions'][i, :] = action
obs = env.step(action)[0]
img = obs['image_observation']
if use_images:
dataset['next_obs'][i, :] = unormalize_image(img)
else:
dataset['next_obs'][i, :] = obs['state_observation']
dataset['next_obs_state'][i, :] = obs['state_observation']
show = True #todo del own
if show:
img = img.reshape(3, imsize, imsize).transpose((1, 2, 0))
img = img[::, :, ::-1]
cv2.imshow('img', img)
cv2.waitKey(1000)
print("keys and shapes:")
for k in dataset.keys():
print(k, dataset[k].shape)
print("done making training data", filename, time.time() - now)
np.save(filename, dataset)
def generate_vae_dataset(variant):
env_class = variant.get('env_class', None)
env_kwargs = variant.get('env_kwargs', None)
env_id = variant.get('env_id', None)
N = variant.get('N', 10000)
test_p = variant.get('test_p', 0.9)
use_cached = variant.get('use_cached', True)
imsize = variant.get('imsize', 84)
num_channels = variant.get('num_channels', 3)
show = variant.get('show', False)
init_camera = variant.get('init_camera', None)
dataset_path = variant.get('dataset_path', None)
oracle_dataset_using_set_to_goal = variant.get(
'oracle_dataset_using_set_to_goal', False)
random_rollout_data = variant.get('random_rollout_data', False)
random_and_oracle_policy_data = variant.get(
'random_and_oracle_policy_data', False)
random_and_oracle_policy_data_split = variant.get(
'random_and_oracle_policy_data_split', 0)
policy_file = variant.get('policy_file', None)
n_random_steps = variant.get('n_random_steps', 100)
vae_dataset_specific_env_kwargs = variant.get(
'vae_dataset_specific_env_kwargs', None)
save_file_prefix = variant.get('save_file_prefix', None)
non_presampled_goal_img_is_garbage = variant.get(
'non_presampled_goal_img_is_garbage', None)
tag = variant.get('tag', '')
from multiworld.core.image_env import ImageEnv, unormalize_image
import rlkit.torch.pytorch_util as ptu
info = {}
if dataset_path is not None:
dataset = load_local_or_remote_file(dataset_path)
N = dataset.shape[0]
else:
if env_kwargs is None:
env_kwargs = {}
if save_file_prefix is None:
save_file_prefix = env_id
if save_file_prefix is None:
save_file_prefix = env_class.__name__
filename = "/tmp/{}_N{}_{}_imsize{}_random_oracle_split_{}{}.npy".format(
save_file_prefix,
str(N),
init_camera.__name__ if init_camera else '',
imsize,
random_and_oracle_policy_data_split,
tag,
)
if use_cached and osp.isfile(filename):
dataset = np.load(filename)
print("loaded data from saved file", filename)
else:
now = time.time()
if env_id is not None:
import gym
import multiworld
multiworld.register_all_envs()
env = gym.make(env_id)
else:
if vae_dataset_specific_env_kwargs is None:
vae_dataset_specific_env_kwargs = {}
for key, val in env_kwargs.items():
if key not in vae_dataset_specific_env_kwargs:
vae_dataset_specific_env_kwargs[key] = val
env = env_class(**vae_dataset_specific_env_kwargs)
if not isinstance(env, ImageEnv):
env = ImageEnv(
env,
imsize,
init_camera=init_camera,
transpose=True,
normalize=True,
non_presampled_goal_img_is_garbage=
non_presampled_goal_img_is_garbage,
)
else:
imsize = env.imsize
env.non_presampled_goal_img_is_garbage = non_presampled_goal_img_is_garbage
env.reset()
info['env'] = env
if random_and_oracle_policy_data:
policy_file = load_local_or_remote_file(policy_file)
policy = policy_file['policy']
policy.to(ptu.device)
if random_rollout_data:
from rlkit.exploration_strategies.ou_strategy import OUStrategy
policy = OUStrategy(env.action_space)
dataset = np.zeros((N, imsize * imsize * num_channels),
dtype=np.uint8)
for i in range(10000):
NP = []
if oracle_dataset_using_set_to_goal:
print(i)
#print('th step')
goal = env.sample_goal()
env.set_to_goal(goal)
obs = env._get_obs()
#img = img.reshape(3, imsize, imsize).transpose()
# img = img[::-1, :, ::-1]
# cv2.imshow('img', img)
# cv2.waitKey(1)
img_1 = obs['image_observation']
NP.append(img_1)
#dataset[i, :] = unormalize_image(img)
img_1 = img_1.reshape(3, imsize, imsize).transpose()
if i % 3 == 0:
cv2.imshow('img1', img_1)
cv2.waitKey(1)
env.reset()
instr = env.generate_new_state(goal)
if i % 3 == 0:
print(instr)
obs = env._get_obs()
# obs = env._get_obs()
img_2 = obs['image_observation']
NP.append(img_2)
NP.append(instr)
img_2 = img_2.reshape(3, imsize, imsize).transpose()
if i % 3 == 0:
cv2.imshow('img2', img_2)
cv2.waitKey(1)
NP = np.array(NP)
print(NP)
idx = str(i)
name = "/home/xiaomin/Downloads/IFIG_DATA_1/" + idx + ".npy"
np.save(open(name, 'wb'), NP)
# radius = input('waiting...')
print("done making training data", filename, time.time() - now)
np.save(filename, dataset)
n = int(N * test_p)
train_dataset = dataset[:n, :]
test_dataset = dataset[n:, :]
return train_dataset, test_dataset, info
